光伏阵列通常被安装在恶劣的室外环境中,因此在运行过程中易发生故障.为了准确识别光伏阵列的故障类型,提出沙猫群优化支持向量机(sand cat swarm optimization support vector machine,SCSO-SVM)用于光伏组件故障识别,且对比支持向量机(support vector machine,SVM)、粒子群优化支持向量机(particle swarm optimized support vector machine,PSO-SVM)、遗传优化支持向量机(genetic optimized support vector machine,GA-SVM)、麻雀优化支持向量机(sparrow optimized support vector ma-chine,SSA-SVM)、灰狼优化支持向量机(gray wolf optimized support vector machine,GWO-SVM)和鲸鱼优化支持向量机(whale optimized support vector machine,WOA-SVM)算法.首先,六种SVM混合算法都克服了SVM诊断结果易受参数初始值影响的缺点,识别精度相较传统SVM算法都有所提升,但是识别时间都增加.其次,7种算法中SCSO-SVM识别效果最好,克服了SVM易受参数初始值的影响,相较SVM识别精度提高了约9.459 4％;是因为更能有效找到SVM惩罚因子和核函数参数.然后,对于同一种算法而言,算法的识别精度是随输入特征减少而降低的,是因为输入特征越少,越不能有效表征光伏组件在不同故障类型下的输出属性.但算法的识别时间却不是随输入特征减少而减短.所以选取合适的输入特征才能兼顾算法的故障识别准确率和效率.最后,发现七种算法的识别效果依赖于数据集的影响.原因可能是各个算法参数选择过多导致泛化性有差异,且依赖参数初始值选择.
Fault Identification of PV Modules Based on SCSO-SVM Algorithm
Since photovoltaic(PV)arrays installationin in harsh outdoor environments,PV faults frequently occur during their operation.To accurately identify the fault types of PV arrays,a sand cat swarm optimization support vector machine(SCSO-SVM)was proposed for PV module fault identification.In addition,the SCSO-SVM,support vector machine(SVM),particle swarm optimized support vector machine(PSO-SVM),genetic optimized support vector machine(GA-SVM),sparrow optimized support vector machine(SSA-SVM),gray wolf optimized support vector machine(GWO-SVM)and whale optimized support vector machine(WOA-SVM)algorithms were compared.First of all,all six SVM hybrid algorithms overcome the disadvantage that SVM diagnosis results were easily affected by the initial values of parameters,and the recognition accuracy was improved compared with traditional SVM algorithms,but the recognition time was increased for all of them.Secondly,SCSO-SVM recognition was the best among the seven algorithms,which overcomed the vulnerability of SVM to the initial values of parameters and improved the recognition accuracy by about 9.459 4％compared to SVM.Because it is more effective in finding the SVM penalty factors and kernel function parameters.Then,for the same algorithm,the recognition accuracy of the algorithm decreased with decreasing input features because the fewer the input features,the less effective it was to characterize the output properties of the PV modules under different fault types.However,the recognition time of the algorithm was not brief with the decrease of the input features.Therefore,the appropriate input features were selected to balance the fault recognition accuracy and efficiency of the algorithm.Finally,it was found that the recognition effect of the seven algorithms depends on the effect of the dataset.The reason may be that there are differences in generalizability due to excessive selection of parameters for each algorithm and dependence on the initial value selection of parameters.
万方数据
维普
